Think of it as the model's working memory for one reply. Everything the model needs to consider has to fit inside it: the system instructions describing how to behave, any source material pulled in to answer the question, the history of the conversation, and the new message. Once the total exceeds the limit, something has to be dropped or summarised, and older parts of the conversation are usually the first to go.
This is why a long chat can start to feel forgetful, and why feeding an entire website into a single prompt is not a workable strategy. Larger context windows help, but they are not a free upgrade. More material in the window can dilute attention, and the relevant sentence can get lost among thousands of irrelevant ones.
The practical answer is selectivity. Rather than pushing everything into the window, a well-built chat widget retrieves only the passages that relate to the question being asked. Clerkzo crawls your site, indexes it, and pulls the relevant sections into each answer, which keeps responses focused on the page that actually addresses the visitor's question instead of a vague blend of your whole site.
Related terms
Browse all 142 terms- Availability WindowAn availability window is a defined block of time during which a service can be booked or a visit will take place, such as between 8am and 12pm. It replaces a precise appointment time with a range you can realistically keep.
- Conversation ContextConversation context is the complete set of information a chatbot has available when composing a reply, including the earlier messages in the conversation and any content retrieved from its knowledge base for the current question.
- Model TemperatureModel temperature is a setting that controls how much randomness an AI model applies when choosing its next word. Lower values produce consistent, predictable output; higher values produce more varied and creative output.
- InferenceInference is the act of running a trained AI model to produce an output from a new input. It is the moment a visitor asks a question and the model generates an answer, as opposed to the earlier training phase that built the model.